Enhancing Flood Inundation Modeling with Memory-Optimized Deep Learning Approaches

Abstract Submission: With advancements in computational technology, simulating flood models with finer-resolution grid cells has become more achievable in recent years. However, high-resolution flood simulations require significant memory resources, often demanding multiple Graphical Processing Units (GPUs), which limits the usability of these models, particularly in time-sensitive scenarios such as storm response. Optimizing memory usage while preserving high-resolution accuracy has thus become a critical challenge. Machine learning models have emerged as a promising solution to this issue by offering faster simulations, although they can also be memory-intensive when trained on large datasets. This study introduces TRITON-Lite, a machine learning surrogate model for TRITON (Two-dimensional Runoff Inundation Toolkit for Operational Needs) flood inundation model. TRITON-Lite not only provides faster and accurate flood inundation predictions but also optimizes memory usage during training. The model was implemented in the Jacks River Watershed and Holly Creek Watershed, both with diverse hydrological and geographical characteristics in the southeastern Georgia. TRITON-Lite was trained on a wide range of flood events at the sub-watershed level in these areas. Testing showed that TRITON-Lite achieved high predictive accuracy while significantly reducing memory requirements by nearly 42% for the Jacks River Watershed and 23.4% for the Holly Creek Watershed. Specifically, TRITON-Lite’s outputs were 83.9% and 75.2% smaller than those generated by TRITON for one flood event, respectively. This approach demonstrates a significant advancement in memory-efficient flood mapping, offering rapid simulations and improving large-scale data handling capabilities.

Learning Objectives/Expected Outcome (Optional) :